fix issues

* Add detection code for avx

* Only check hardware when option is ON

* Modify per code review sugguestions

* Build locally will detect CPU

* Fixes CMake style to use lowercase like everywhere else

* cleanup

* fix merge

* linux/gcc version for testing

* msvc combines avx2 and fma into /arch:AVX2 so check for both

* cleanup

* msvc only version

* style

* Update FindSIMD.cmake

---------

Co-authored-by: Howard Su <howard0su@gmail.com>
Co-authored-by: Jeremy Dunn <jeremydunn123@gmail.com>

.github/workflows/build.yml

@@ -288,6 +288,7 @@ jobs:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
+      SDE_VERSION: 9.21.1-2023-04-24
 
     strategy:
       matrix:
@@ -383,11 +384,23 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
+        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
         run: |
           cd build
           ctest -C Release --verbose --timeout 900
 
+      - name: Test (Intel SDE)
+        id: cmake_test_sde
+        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz"
+          # for some weird reason windows tar doesn't like sde tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
+          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
+          $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
+          cd build
+          & $sde -future -- ctest -C Release --verbose --timeout 900
+
       - name: Determine tag name
         id: tag
         shell: bash

CMakeLists.txt

@@ -10,7 +10,7 @@ endif()
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
     # configure project version
@@ -44,7 +44,7 @@ endif()
 
 # general
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
-option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      OFF)
+option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
 
 # debug
@@ -510,6 +510,10 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
     message(STATUS "x86 detected")
     if (MSVC)
+        # instruction set detection for MSVC only
+        if (LLAMA_NATIVE)
+            include(cmake/FindSIMD.cmake)
+        endif ()
         if (LLAMA_AVX512)
             add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
             add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)

cmake/FindSIMD.cmake

@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(LLAMA_AVX OFF)
+else()
+    set(LLAMA_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(LLAMA_AVX2 OFF)
+else()
+    set(LLAMA_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(LLAMA_AVX512 OFF)
+else()
+    set(LLAMA_AVX512 ON)
+endif()

ggml-cuda.cu

@@ -6893,6 +6893,8 @@ static void ggml_cuda_op_mul_mat(
     int64_t  row_low[GGML_CUDA_MAX_DEVICES];
     int64_t row_high[GGML_CUDA_MAX_DEVICES];
 
+    int used_devices = 0;
+
     for (int64_t id = 0; id < g_device_count; ++id) {
         // by default, use all rows
         row_low[id]  = 0;
@@ -6920,6 +6922,8 @@ static void ggml_cuda_op_mul_mat(
             continue;
         }
 
+        used_devices++;
+
         const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
         const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
 
@@ -6958,12 +6962,12 @@ static void ggml_cuda_op_mul_mat(
 
     // if multiple devices are used they need to wait for the main device
     // here an event is recorded that signals that the main device has finished calculating the input data
-    if (split && g_device_count > 1) {
+    if (split && used_devices > 1) {
         CUDA_CHECK(ggml_cuda_set_device(g_main_device));
         CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0]));
     }
 
-    const int64_t src1_col_stride = split && g_device_count > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
+    const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
     for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
         const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0;
         const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
@@ -7079,6 +7083,9 @@ static void ggml_cuda_op_mul_mat(
     }
 
     for (int64_t id = 0; id < g_device_count; ++id) {
+        if ((!split && id != g_main_device) || row_low[id] == row_high[id]) {
+            continue;
+        }
         CUDA_CHECK(ggml_cuda_set_device(id));
 
         // free buffers again when done
@@ -7103,6 +7110,9 @@ static void ggml_cuda_op_mul_mat(
 
         CUDA_CHECK(ggml_cuda_set_device(g_main_device));
         for (int64_t id = 0; id < g_device_count; ++id) {
+            if (row_low[id] == row_high[id]) {
+                continue;
+            }
             for (int64_t is = 0; is < is_max; ++is) {
                 CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
             }
@@ -7400,7 +7410,7 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const
 
 static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool all_on_device =
-        (src0->backend == GGML_BACKEND_GPU) &&
+        (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
         (src1->backend == GGML_BACKEND_GPU) &&
         ( dst->backend == GGML_BACKEND_GPU);